Inefficient use of Vec::push in prepare_uniform_components function

[CrazyRoka]

I created a stress test scenario with the Bevy engine (link) where I created 100K entities. I noticed that the prepare_uniform_components function was using 4.23% of the CPU time, and upon investigating the code, I found that the Vec3::push method was being called for each of the entities, which is not efficient.

To improve performance, I suggest copying all the objects in batches instead of pushing them one by one. Apart from that, DynamicUniformBuffer::write can be optimized as well. I am attaching my Flamegraph to this issue.

bevy/crates/bevy_render/src/extract_component.rs

Lines 126 to 153 in aefe1f0

	fn prepare_uniform_components<C: Component>(
	mut commands: Commands,
	render_device: Res<RenderDevice>,
	render_queue: Res<RenderQueue>,
	mut component_uniforms: ResMut<ComponentUniforms<C>>,
	components: Query<(Entity, &C)>,
	) where
	C: ShaderType + WriteInto + Clone,
	{
	component_uniforms.uniforms.clear();
	let entities = components
	.iter()
	.map(|(entity, component)| {
	(
	entity,
	DynamicUniformIndex::<C> {
	index: component_uniforms.uniforms.push(component.clone()),
	marker: PhantomData,
	},
	)
	})
	.collect::<Vec<_>>();
	commands.insert_or_spawn_batch(entities);
	component_uniforms
	.uniforms
	.write_buffer(&render_device, &render_queue);
	}

[JMS55]

You may be interested in weighing in on #8204

[james7132]

@CrazyRoka turns out the values Vec is basically unused and was causing unnecessary copies and allocations. #8299 removes it. As for the performance issues with encase, I strongly suggest making a PR upstream to address it. The flamegraph you provided definitely shows there's some non-trivial overhead when using it.

[CrazyRoka]

@james7132 Thank you for your PR to improve Bevy performance. I have 2 more questions:

1. How did you measure performance improvements in your PR?
2. Can we replace component_uniforms.uniforms.push(component.clone()) with batch write? I was diving deeper into the implementation and it looks like each push call creates Writer Cursor, writes and closes this complex object. I think it will make a difference to open WriteCursor once and write all the components together.

[james7132]

1. How did you measure performance improvements in your PR?

I used Tracy and the trace_tracy feature. You can find the docs on how to use it under docs/profiling.md in the repo.

2. Can we replace component_uniforms.uniforms.push(component.clone()) with batch write? I was diving deeper into the implementation and it looks like each push call creates Writer Cursor, writes and closes this complex object. I think it will make a difference to open WriteCursor once and write all the components together.

Batch insertion probably is for the best here. We need to construct both the buffer and the Vec for Commands, and preallocating then skipping the capacity checks likely will have a significant speedup. Though with that said, with dynamic uniform offsets being pessimistically aligned to 256 bytes, we might not see much gain there.